August 15, 2026 — The Guardian reported that secondhand booksellers in the UK and Ireland have seen “strange” bulk buys that they suspect are headed to AI labs. The surge in AI bulk book orders follows earlier reporting that Anthropic spent millions acquiring books to scan for “data acquisition,” according to The Guardian.
The pattern behind AI bulk book orders
Booksellers say the orders stand out. They arrive in clusters, target specific titles, and move fast. That behavior aligns with how teams assemble large text corpora: identify valuable domains, sweep inventory, and ship quickly for processing. The net effect is a quiet, physical supply chain for training data built from used stock rather than publisher deals. The Guardian’s account points to a procurement shift as labs chase long-form, rights‑encumbered text at scale.
Why secondhand? Price, variety, and a lower profile. Backlist titles and technical books can be cheaper and easier to source in bulk through used channels. The tactic also disperses buying across many sellers, which reduces attention until patterns appear. None of that changes the legal stakes, but it does change where the pressure lands: on indie shops and charity outlets suddenly fielding large, unusual requests.
What secondhand buying signals about AI training
The orders tell us something about model needs. Long, well‑edited prose, niche nonfiction, and specialist manuals remain prized for pretraining. Web text alone isn’t enough. When a lab sweeps hundreds or thousands of volumes, it’s seeking density: fewer duplicates, higher editorial quality, and cleaner metadata. That aim matches the “data acquisition” push The Guardian tied to Anthropic and suggests this isn’t a one‑off. It’s a method.
For booksellers, that demand can distort local markets. A few targeted purchases can empty shelves in key categories and nudge prices up for real readers. Libraries and universities may face similar pressures if resellers start funneling discards to buyers focused on ingestion rather than readership. The winners are middlemen who spot the pattern early; the losers are patrons who find gaps where a section used to be.
Does buying used dodge copyright? Short answer: no
Owning a book gives the right to resell it. It does not grant the right to reproduce it. In the UK, commercial text and data mining requires permission unless an explicit exception applies. The government’s own guidance says the text‑and‑data‑analysis exception covers non‑commercial research only; copying for commercial mining needs a licence (UK Intellectual Property Office). Buying secondhand changes the seller, not the law.
Ireland, within the EU, sits under the bloc’s text and data mining regime. Article 4 of the Copyright in the Digital Single Market Directive permits mining for commercial uses unless rights holders have opted out in a machine‑readable way. If a publisher has opted out, mining is off‑limits without a licence (European Commission). Again, the provenance of the physical copy doesn’t erase reproduction rights.
Litigation pressure is building in parallel. In the United States, the Authors Guild is pursuing a class action against OpenAI over alleged unlicensed use of books in training. That case, while under different legal standards, shows where this fight is heading: toward paid, auditable access rather than “find it and scan it” shortcuts (Authors Guild).
Why this matters for sellers, publishers, and labs
The Guardian’s reporting gives independent shops a heads‑up. Large, sudden buys can be legitimate, but patterns matter. If requests target specific ISBNs across multiple stores, arrive via newly formed accounts, and demand rapid courier pickup, the end use is likely dataset assembly. Sellers should decide their policy now. Some will prefer to take the sale; others may cap quantities or require business details for bulk orders.
Publishers get a signal too. If used channels are being tapped at scale, it suggests current licensing offers aren’t meeting buyer demand on price, breadth, or speed. That’s an opening to propose standardized, tiered licences that cover specific uses, retention periods, and audit rights. The UK Publishers Association has argued for clear rules and paid access; this is a chance to put those principles into workable deals (Publishers Association).
For labs, the risk profile is obvious. Scanning a lawfully acquired book still creates a copy. If the use isn’t covered by an exception or licence, the exposure sits squarely on the buyer, not the bookseller. Beyond legal risk, covert procurement erodes trust at a time when regulators are already asking for training data transparency.
What to watch next as AI bulk book orders rise
Expect more scrutiny of AI bulk book orders by trade groups and regulators. In the UK, that could mean guidance to retailers on handling large requests tied to data mining. In the EU, rights‑holder opt‑outs will be tested as buyers weigh the cost of licences against the temptation of the used market. Either way, the technical value of curated print is now clear, and so are the stakes.
Practical steps are straightforward. Shops can keep basic records of bulk requests, set quantity limits on flagged titles, and share patterns with their trade associations. Publishers can expand machine‑readable opt‑outs and publish licensing terms that are easy to find and easy to buy. Labs that want to avoid litigation and reputational fallout can move first: disclose sources at a high level, license what’s needed, and stop treating book stock as a loophole.
The Guardian has surfaced a telling signal. The hunger for quality text hasn’t gone away; it’s migrated to a quieter channel. How the industry responds will decide whether that channel becomes a bridge to licensed access or a legal cul‑de‑sac for the next wave of AI bulk book orders. For more on this, see anthropic.com and reuters.com and bloomberg.com.
Related reading: AI in Education • Data Privacy • AI in Society